1 /*
2 * Copyright (c) 2018, Alliance for Open Media. All rights reserved.
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #include "warp_plane_neon.h"
13
horizontal_filter_4x1_f4(const uint8x16_t in,int sx,int alpha)14 static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f4(const uint8x16_t in,
15 int sx, int alpha) {
16 const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1));
17
18 // Loading the 8 filter taps
19 int16x8_t f[4];
20 load_filters_4(f, sx, alpha);
21
22 int16x8_t in16_lo = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(in)));
23 int16x8_t in16_hi = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(in)));
24
25 int16x8_t m0 = vmulq_s16(f[0], in16_lo);
26 int16x8_t m1 = vmulq_s16(f[1], vextq_s16(in16_lo, in16_hi, 1));
27 int16x8_t m2 = vmulq_s16(f[2], vextq_s16(in16_lo, in16_hi, 2));
28 int16x8_t m3 = vmulq_s16(f[3], vextq_s16(in16_lo, in16_hi, 3));
29
30 int32x4_t m0123_pairs[] = { vpaddlq_s16(m0), vpaddlq_s16(m1), vpaddlq_s16(m2),
31 vpaddlq_s16(m3) };
32
33 int32x4_t tmp_res_low = horizontal_add_4d_s32x4(m0123_pairs);
34
35 tmp_res_low = vaddq_s32(tmp_res_low, add_const);
36
37 uint16x8_t res =
38 vcombine_u16(vqrshrun_n_s32(tmp_res_low, ROUND0_BITS), vdup_n_u16(0));
39 return vreinterpretq_s16_u16(res);
40 }
41
horizontal_filter_8x1_f8(const uint8x16_t in,int sx,int alpha)42 static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f8(const uint8x16_t in,
43 int sx, int alpha) {
44 const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1));
45
46 // Loading the 8 filter taps
47 int16x8_t f[8];
48 load_filters_8(f, sx, alpha);
49
50 int16x8_t in16_lo = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(in)));
51 int16x8_t in16_hi = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(in)));
52
53 int16x8_t m0 = vmulq_s16(f[0], in16_lo);
54 int16x8_t m1 = vmulq_s16(f[1], vextq_s16(in16_lo, in16_hi, 1));
55 int16x8_t m2 = vmulq_s16(f[2], vextq_s16(in16_lo, in16_hi, 2));
56 int16x8_t m3 = vmulq_s16(f[3], vextq_s16(in16_lo, in16_hi, 3));
57 int16x8_t m4 = vmulq_s16(f[4], vextq_s16(in16_lo, in16_hi, 4));
58 int16x8_t m5 = vmulq_s16(f[5], vextq_s16(in16_lo, in16_hi, 5));
59 int16x8_t m6 = vmulq_s16(f[6], vextq_s16(in16_lo, in16_hi, 6));
60 int16x8_t m7 = vmulq_s16(f[7], vextq_s16(in16_lo, in16_hi, 7));
61
62 int32x4_t m0123_pairs[] = { vpaddlq_s16(m0), vpaddlq_s16(m1), vpaddlq_s16(m2),
63 vpaddlq_s16(m3) };
64 int32x4_t m4567_pairs[] = { vpaddlq_s16(m4), vpaddlq_s16(m5), vpaddlq_s16(m6),
65 vpaddlq_s16(m7) };
66
67 int32x4_t tmp_res_low = horizontal_add_4d_s32x4(m0123_pairs);
68 int32x4_t tmp_res_high = horizontal_add_4d_s32x4(m4567_pairs);
69
70 tmp_res_low = vaddq_s32(tmp_res_low, add_const);
71 tmp_res_high = vaddq_s32(tmp_res_high, add_const);
72
73 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(tmp_res_low, ROUND0_BITS),
74 vqrshrun_n_s32(tmp_res_high, ROUND0_BITS));
75 return vreinterpretq_s16_u16(res);
76 }
77
78 static AOM_FORCE_INLINE int16x8_t
horizontal_filter_4x1_f1_beta0(const uint8x16_t in,int16x8_t f_s16)79 horizontal_filter_4x1_f1_beta0(const uint8x16_t in, int16x8_t f_s16) {
80 const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1));
81
82 int16x8_t in16_lo = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(in)));
83 int16x8_t in16_hi = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(in)));
84
85 int16x8_t m0 = vmulq_s16(f_s16, in16_lo);
86 int16x8_t m1 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 1));
87 int16x8_t m2 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 2));
88 int16x8_t m3 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 3));
89
90 int32x4_t m0123_pairs[] = { vpaddlq_s16(m0), vpaddlq_s16(m1), vpaddlq_s16(m2),
91 vpaddlq_s16(m3) };
92
93 int32x4_t tmp_res_low = horizontal_add_4d_s32x4(m0123_pairs);
94
95 tmp_res_low = vaddq_s32(tmp_res_low, add_const);
96
97 uint16x8_t res =
98 vcombine_u16(vqrshrun_n_s32(tmp_res_low, ROUND0_BITS), vdup_n_u16(0));
99 return vreinterpretq_s16_u16(res);
100 }
101
horizontal_filter_4x1_f1(const uint8x16_t in,int sx)102 static AOM_FORCE_INLINE int16x8_t horizontal_filter_4x1_f1(const uint8x16_t in,
103 int sx) {
104 int16x8_t f_s16 =
105 vld1q_s16((int16_t *)(av1_warped_filter + (sx >> WARPEDDIFF_PREC_BITS)));
106 return horizontal_filter_4x1_f1_beta0(in, f_s16);
107 }
108
109 static AOM_FORCE_INLINE int16x8_t
horizontal_filter_8x1_f1_beta0(const uint8x16_t in,int16x8_t f_s16)110 horizontal_filter_8x1_f1_beta0(const uint8x16_t in, int16x8_t f_s16) {
111 const int32x4_t add_const = vdupq_n_s32(1 << (8 + FILTER_BITS - 1));
112
113 int16x8_t in16_lo = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(in)));
114 int16x8_t in16_hi = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(in)));
115
116 int16x8_t m0 = vmulq_s16(f_s16, in16_lo);
117 int16x8_t m1 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 1));
118 int16x8_t m2 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 2));
119 int16x8_t m3 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 3));
120 int16x8_t m4 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 4));
121 int16x8_t m5 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 5));
122 int16x8_t m6 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 6));
123 int16x8_t m7 = vmulq_s16(f_s16, vextq_s16(in16_lo, in16_hi, 7));
124
125 int32x4_t m0123_pairs[] = { vpaddlq_s16(m0), vpaddlq_s16(m1), vpaddlq_s16(m2),
126 vpaddlq_s16(m3) };
127 int32x4_t m4567_pairs[] = { vpaddlq_s16(m4), vpaddlq_s16(m5), vpaddlq_s16(m6),
128 vpaddlq_s16(m7) };
129
130 int32x4_t tmp_res_low = horizontal_add_4d_s32x4(m0123_pairs);
131 int32x4_t tmp_res_high = horizontal_add_4d_s32x4(m4567_pairs);
132
133 tmp_res_low = vaddq_s32(tmp_res_low, add_const);
134 tmp_res_high = vaddq_s32(tmp_res_high, add_const);
135
136 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(tmp_res_low, ROUND0_BITS),
137 vqrshrun_n_s32(tmp_res_high, ROUND0_BITS));
138 return vreinterpretq_s16_u16(res);
139 }
140
horizontal_filter_8x1_f1(const uint8x16_t in,int sx)141 static AOM_FORCE_INLINE int16x8_t horizontal_filter_8x1_f1(const uint8x16_t in,
142 int sx) {
143 int16x8_t f_s16 =
144 vld1q_s16((int16_t *)(av1_warped_filter + (sx >> WARPEDDIFF_PREC_BITS)));
145 return horizontal_filter_8x1_f1_beta0(in, f_s16);
146 }
147
vertical_filter_4x1_f1(const int16x8_t * src,int32x4_t * res,int sy)148 static AOM_FORCE_INLINE void vertical_filter_4x1_f1(const int16x8_t *src,
149 int32x4_t *res, int sy) {
150 int16x4_t s0 = vget_low_s16(src[0]);
151 int16x4_t s1 = vget_low_s16(src[1]);
152 int16x4_t s2 = vget_low_s16(src[2]);
153 int16x4_t s3 = vget_low_s16(src[3]);
154 int16x4_t s4 = vget_low_s16(src[4]);
155 int16x4_t s5 = vget_low_s16(src[5]);
156 int16x4_t s6 = vget_low_s16(src[6]);
157 int16x4_t s7 = vget_low_s16(src[7]);
158
159 int16x8_t f =
160 vld1q_s16((int16_t *)(av1_warped_filter + (sy >> WARPEDDIFF_PREC_BITS)));
161
162 int32x4_t m0123 = vmull_lane_s16(s0, vget_low_s16(f), 0);
163 m0123 = vmlal_lane_s16(m0123, s1, vget_low_s16(f), 1);
164 m0123 = vmlal_lane_s16(m0123, s2, vget_low_s16(f), 2);
165 m0123 = vmlal_lane_s16(m0123, s3, vget_low_s16(f), 3);
166 m0123 = vmlal_lane_s16(m0123, s4, vget_high_s16(f), 0);
167 m0123 = vmlal_lane_s16(m0123, s5, vget_high_s16(f), 1);
168 m0123 = vmlal_lane_s16(m0123, s6, vget_high_s16(f), 2);
169 m0123 = vmlal_lane_s16(m0123, s7, vget_high_s16(f), 3);
170
171 *res = m0123;
172 }
173
vertical_filter_4x1_f4(const int16x8_t * src,int32x4_t * res,int sy,int gamma)174 static AOM_FORCE_INLINE void vertical_filter_4x1_f4(const int16x8_t *src,
175 int32x4_t *res, int sy,
176 int gamma) {
177 int16x8_t s0, s1, s2, s3;
178 transpose_elems_s16_4x8(
179 vget_low_s16(src[0]), vget_low_s16(src[1]), vget_low_s16(src[2]),
180 vget_low_s16(src[3]), vget_low_s16(src[4]), vget_low_s16(src[5]),
181 vget_low_s16(src[6]), vget_low_s16(src[7]), &s0, &s1, &s2, &s3);
182
183 int16x8_t f[4];
184 load_filters_4(f, sy, gamma);
185
186 int32x4_t m0 = vmull_s16(vget_low_s16(s0), vget_low_s16(f[0]));
187 m0 = vmlal_s16(m0, vget_high_s16(s0), vget_high_s16(f[0]));
188 int32x4_t m1 = vmull_s16(vget_low_s16(s1), vget_low_s16(f[1]));
189 m1 = vmlal_s16(m1, vget_high_s16(s1), vget_high_s16(f[1]));
190 int32x4_t m2 = vmull_s16(vget_low_s16(s2), vget_low_s16(f[2]));
191 m2 = vmlal_s16(m2, vget_high_s16(s2), vget_high_s16(f[2]));
192 int32x4_t m3 = vmull_s16(vget_low_s16(s3), vget_low_s16(f[3]));
193 m3 = vmlal_s16(m3, vget_high_s16(s3), vget_high_s16(f[3]));
194
195 int32x4_t m0123_pairs[] = { m0, m1, m2, m3 };
196
197 *res = horizontal_add_4d_s32x4(m0123_pairs);
198 }
199
vertical_filter_8x1_f1(const int16x8_t * src,int32x4_t * res_low,int32x4_t * res_high,int sy)200 static AOM_FORCE_INLINE void vertical_filter_8x1_f1(const int16x8_t *src,
201 int32x4_t *res_low,
202 int32x4_t *res_high,
203 int sy) {
204 int16x8_t s0 = src[0];
205 int16x8_t s1 = src[1];
206 int16x8_t s2 = src[2];
207 int16x8_t s3 = src[3];
208 int16x8_t s4 = src[4];
209 int16x8_t s5 = src[5];
210 int16x8_t s6 = src[6];
211 int16x8_t s7 = src[7];
212
213 int16x8_t f =
214 vld1q_s16((int16_t *)(av1_warped_filter + (sy >> WARPEDDIFF_PREC_BITS)));
215
216 int32x4_t m0123 = vmull_lane_s16(vget_low_s16(s0), vget_low_s16(f), 0);
217 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s1), vget_low_s16(f), 1);
218 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s2), vget_low_s16(f), 2);
219 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s3), vget_low_s16(f), 3);
220 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s4), vget_high_s16(f), 0);
221 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s5), vget_high_s16(f), 1);
222 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s6), vget_high_s16(f), 2);
223 m0123 = vmlal_lane_s16(m0123, vget_low_s16(s7), vget_high_s16(f), 3);
224
225 int32x4_t m4567 = vmull_lane_s16(vget_high_s16(s0), vget_low_s16(f), 0);
226 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s1), vget_low_s16(f), 1);
227 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s2), vget_low_s16(f), 2);
228 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s3), vget_low_s16(f), 3);
229 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s4), vget_high_s16(f), 0);
230 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s5), vget_high_s16(f), 1);
231 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s6), vget_high_s16(f), 2);
232 m4567 = vmlal_lane_s16(m4567, vget_high_s16(s7), vget_high_s16(f), 3);
233
234 *res_low = m0123;
235 *res_high = m4567;
236 }
237
vertical_filter_8x1_f8(const int16x8_t * src,int32x4_t * res_low,int32x4_t * res_high,int sy,int gamma)238 static AOM_FORCE_INLINE void vertical_filter_8x1_f8(const int16x8_t *src,
239 int32x4_t *res_low,
240 int32x4_t *res_high, int sy,
241 int gamma) {
242 int16x8_t s0 = src[0];
243 int16x8_t s1 = src[1];
244 int16x8_t s2 = src[2];
245 int16x8_t s3 = src[3];
246 int16x8_t s4 = src[4];
247 int16x8_t s5 = src[5];
248 int16x8_t s6 = src[6];
249 int16x8_t s7 = src[7];
250 transpose_elems_inplace_s16_8x8(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
251
252 int16x8_t f[8];
253 load_filters_8(f, sy, gamma);
254
255 int32x4_t m0 = vmull_s16(vget_low_s16(s0), vget_low_s16(f[0]));
256 m0 = vmlal_s16(m0, vget_high_s16(s0), vget_high_s16(f[0]));
257 int32x4_t m1 = vmull_s16(vget_low_s16(s1), vget_low_s16(f[1]));
258 m1 = vmlal_s16(m1, vget_high_s16(s1), vget_high_s16(f[1]));
259 int32x4_t m2 = vmull_s16(vget_low_s16(s2), vget_low_s16(f[2]));
260 m2 = vmlal_s16(m2, vget_high_s16(s2), vget_high_s16(f[2]));
261 int32x4_t m3 = vmull_s16(vget_low_s16(s3), vget_low_s16(f[3]));
262 m3 = vmlal_s16(m3, vget_high_s16(s3), vget_high_s16(f[3]));
263 int32x4_t m4 = vmull_s16(vget_low_s16(s4), vget_low_s16(f[4]));
264 m4 = vmlal_s16(m4, vget_high_s16(s4), vget_high_s16(f[4]));
265 int32x4_t m5 = vmull_s16(vget_low_s16(s5), vget_low_s16(f[5]));
266 m5 = vmlal_s16(m5, vget_high_s16(s5), vget_high_s16(f[5]));
267 int32x4_t m6 = vmull_s16(vget_low_s16(s6), vget_low_s16(f[6]));
268 m6 = vmlal_s16(m6, vget_high_s16(s6), vget_high_s16(f[6]));
269 int32x4_t m7 = vmull_s16(vget_low_s16(s7), vget_low_s16(f[7]));
270 m7 = vmlal_s16(m7, vget_high_s16(s7), vget_high_s16(f[7]));
271
272 int32x4_t m0123_pairs[] = { m0, m1, m2, m3 };
273 int32x4_t m4567_pairs[] = { m4, m5, m6, m7 };
274
275 *res_low = horizontal_add_4d_s32x4(m0123_pairs);
276 *res_high = horizontal_add_4d_s32x4(m4567_pairs);
277 }
278
av1_warp_affine_neon(const int32_t * mat,const uint8_t * ref,int width,int height,int stride,uint8_t * pred,int p_col,int p_row,int p_width,int p_height,int p_stride,int subsampling_x,int subsampling_y,ConvolveParams * conv_params,int16_t alpha,int16_t beta,int16_t gamma,int16_t delta)279 void av1_warp_affine_neon(const int32_t *mat, const uint8_t *ref, int width,
280 int height, int stride, uint8_t *pred, int p_col,
281 int p_row, int p_width, int p_height, int p_stride,
282 int subsampling_x, int subsampling_y,
283 ConvolveParams *conv_params, int16_t alpha,
284 int16_t beta, int16_t gamma, int16_t delta) {
285 av1_warp_affine_common(mat, ref, width, height, stride, pred, p_col, p_row,
286 p_width, p_height, p_stride, subsampling_x,
287 subsampling_y, conv_params, alpha, beta, gamma, delta);
288 }
289